Ai2 "LoadError: Numerically Inconsistent" #137
Comments
|
Do you have more of the stacktrace? The first line above is already from CDDLib, which is beyond our paygrade! |
|
I tried but am now not able to reproduce it with the same setup. I wonder if something with merging in master fixed it but it took restarting my environment to see that change? I am using Revise so I wouldn't have expected that and I had tried after my most recent merge. |
|
That may be because you were using the algorithm that now is called |
|
Thanks for the suggestion! I tried running with Ai2h() instead of Ai2() in master and it just runs indefinitely (instead of leading to an error). As a result, I'm still not sure what changed to make the old error disappear. |
|
The errors are perhaps related to how the inputs are used in the algorithm? Note that both Ai2h resp. the Ai2z approaches as shown below work fine in the example: input_set = HPolytope([HalfSpace([1.0, 0.0, 0.0, 0.0], 0.3845579517032085),
HalfSpace([0.0, 1.0, 0.0, 0.0], 0.8331193586918577),
HalfSpace([0.0, 0.0, 1.0, 0.0], 0.6607352829885846),
HalfSpace([0.0, 0.0, 0.0, 1.0], 0.270232549769154),
HalfSpace([-1.0, 0.0, 0.0, 0.0], 0.022984248429274823),
HalfSpace([0.0, -1.0, 0.0, 0.0], 1.061033037362016),
HalfSpace([0.0, 0.0, -1.0, 0.0], 0.734914583454296),
HalfSpace([0.0, 0.0, 0.0, -1.0], 0.7494557834041424)]);
# "Ai2h"
@time begin
H1 = overapproximate(network.layers[1].weights * input_set + network.layers[1].bias, Hyperrectangle)
H1 = rectify(H1)
H2 = overapproximate(network.layers[2].weights * H1 + network.layers[2].bias, Hyperrectangle);
end
0.006664 seconds (1.06 k allocations: 78.672 KiB)
Hyperrectangle{Float64,Array{Float64,1},Array{Float64,1}}([2.3950844668146862, -0.13751415227951902, -2.1096209129694556, -0.6299855269624741], [3.1193665559469745, 2.295359657778874, 2.454380545304072, 2.4271559068135793])
# "Ai2z"
@time begin
input_set = overapproximate(input_set, Hyperrectangle) # exact
input_set = convert(Zonotope, input_set) # also exact
Z1 = affine_map(network.layers[1].weights, input_set, network.layers[1].bias)
Z1 = overapproximate(Rectification(Z1), Zonotope)
Z2 = affine_map(network.layers[2].weights, Z1, network.layers[2].bias)
end
0.009373 seconds (536 allocations: 42.047 KiB)
Zonotope{Float64,Array{Float64,1},Array{Float64,2}}([1.7115830883645933, -0.23495075807030685, -2.11436002496305, -0.3868240216039214], [-0.32634300914634634 0.7369189071413871 … 0.18472489913317133 0.15586653347355545; 0.2154030641984947 0.5327045705827566 … -0.1545206676442892 0.015391252910196256; 0.038867901587761676 -0.5260435688205888 … -0.09919108055920038 0.1201795609033839; -0.20050663784682085 -0.44243336453491733 … 0.1539373002888789 -0.23997320951003187])plot(project(H2, [1, 2]), lab="H2")
plot!(project(Z2, [1, 2]), lab="Z2")
|
|
@mforets what you call |
ok, thanks.. i thought the out_Box = forward_network(Box(), network, input_set)
plot!(project(out_Box, [1, 2]), lab="Box")
out_Ai2z = forward_network(Ai2z(), network, input_set)
plot!(project(out_Ai2z, [1, 2]), lab="Ai2z")
|
|
I also saw the call with Ai2h hang forever. When I interrupted eventually, I remember the stacktrace went through |
where/why is |
|
It is used in the old implementation, |
|
@mforets the h in |
With network:
Results in the following output and stacktrace:
The text was updated successfully, but these errors were encountered: